Stronger independent response evaluation, 2026-10-12¶

Current scope, 13 October¶

The original deployment and historical resource limits below are retained for provenance. The current response campaign has five retained workers, ten deliberately truncated archived-view Coins searches, and no admission controller. It is not necessary to finish all fifteen original searches to finish the revised paper.

At 15:22 UTC the Gamma/Coins owner relayed Omar's instruction to use stronger response searches only where they can change a reported equilibrium conclusion. Full-view Coins replaces the local-view main configuration; therefore the remaining local-view searches are no longer a required comparison. An exact best-response result, infeasibility proof, or an independently validated profitable deviation can resolve a negative claim without exhaustive additional response training. An unsuccessful finite search still cannot establish an unrestricted equilibrium.

At 15:48 UTC W1/W3 checked the source, controller arguments, scheduler owner, working directory, output paths and all fifteen worker identities under the shared parent admission lock. It stopped admission controller 3700144 and cancelled only these ten archived local-view Coins workers:

All 193 saved checkpoints, logs, protocol/runtime files and original input archives remain in place. None of these workers had a final held-out result.json when retired. Do not interpret a complete set of training checkpoints as completed validation. Original campaign/deployment/PID bytes were preserved before mutation, and the original fifteen-worker plan remains in the ledger. Records use truncated_by_scope_change, not a numerical or execution failure. These incomplete searches yield no stability certificate.

The remote root's selective-retirement.json contains the full artifact inventory; the compact durable record is results/strong-response-selective-retirement-2026-10-13.json. Slurm accounting independently reports all ten jobs cancelled, and a live queue check found all ten absent. The response allowance decreases from fifteen to five, freeing ten physical-worker slots (twenty allocated Slurm CPUs). The Gamma/Coins owner was notified that these slots can be transferred to its full-budget confirmation wave, subject to the unchanged overall 83-worker maximum and global 177/181/185 admission ceilings. A message offering the transfer is distinct from that owner's completed redeployment. No baseline training budgets, scientific source pins, thresholds, other owners' jobs or external-study jobs changed.

Territory/IPPO jobs 61989--61993 remain running from cec65cf, initially with five or six saved candidates each. Scheduler-tracking process 3947588 holds the root lock, tracks only those retained jobs and has no submission or cancellation code. Once they finish the root becomes complete_with_retirements, rather than claiming all fifteen searches were completed. The local fifteen-minute mailbox watcher recognizes that state and the retired inventory; its original deduplication ledger is preserved. The separate Territory/IPPO-CER searches 61999 and 62000 remain within the Pursuit campaign's existing allowance and require the same usefulness assessment. No new full-view Coins response arm has been admitted; any future arm must use fully_observable=True in every environment factory and use independently evaluated responses only where the conclusion remains unresolved. These ten transferred slots must not also be reused for that arm.

The paper target relayed by the owner is Thursday 15 October midday UK. The older all-search completion estimate is superseded by this selective scope. PR2 repair remains with the CER owner; the completed ten-profile exact/search calibration and the completed fixed-source Pursuit continuation cohort still require final evidence integration.

Omar authorized stronger response evaluation on cpucluster, coordination with the CER and fresh-rerun agents, and corresponding paper updates. This study evaluates fixed archived policies; it does not retrain the original joint profiles or relax their thresholds or epsilon 0.05.

Fixed protocol¶

The fifteen original seed profiles are Coins IPPO and IQL-CER (five seeds each) and Territory IPPO (five seeds). Coins uses the archived local-view factory, explicitly fully_observable=False. The other agent's fresh full-view Coins replacements are a separate information condition and cannot replace these inputs. Monitor numbering is recovered from each archived profile's recorded permutation where necessary.

For each player, opponents are frozen. Independent PPO and IQL searches each have three fresh starts; the matching learner additionally has a native warm start with fresh optimization state. Spatial fresh IQL uses its recurrent architecture, while the small exact calibration uses MLP-IQL so that its opponents can be scored by the public-Markov oracle. Warm starts preserve their original architecture. Responses optimize the undiscounted terminal temporal utility, with CER disabled and its update mode reset to uncorrected. Fresh responses receive the public elapsed-action clock.

Training retains the original declared 50k/200k checkpoints, rounded to 50,176/200,192 actions. Increasing response-validation episodes from 512 to 4096 addresses the saved calibration's demonstrated candidate-selection noise. This is a prospective protocol upgrade, not a re-selection using the preserved calibration's held-out outcomes. Calibration itself retains its original 512/1000 validation/test protocol.

Selection uses 4096 episodes starting at seed * 1_000_000 + 700503. The held-out evaluation uses 8192 disjoint episodes starting at seed * 1_000_000 + 800503, after candidate selection is finished. Baseline and deviating profiles use paired environment seeds. The original policy is always an available response. Raw paired success arrays, checkpoint hashes, source revisions, package versions, candidate choices and optimization budgets remain available for audit.

Report maximum found unilateral gain separately from summed NashConv. The paired exact-binomial assessment allocates delta 0.025 and corrects over agents and both response budgets across all ten Coins or five Territory profiles. Satisfaction/payoff-ceiling assessment has its own delta 0.025. Intervals cover rollout error for the selected deviations only. They can establish a profitable-deviation witness; small found gains and search passes do not give upper bounds on true exploitability. Formal certificates require the separate payoff-ceiling bound or a valid exact best-response oracle.

Independent admission and preserved cohorts¶

The old 75-worker spatial controller had submitted no workers. Its single pilot gate made response evaluation wait for four unrelated Commons training controls. Separate the already authorized plan into:

The combined allowance remains eight, with the shared admission lock and total-user cap 181. The separately owned 42-worker Pursuit controller remains at four. No scheduler worker is cancelled or altered. Preserve the original empty 75-worker ledger and its handoff metadata; only its lightweight controller is superseded after verifying that it still has zero jobs.

The shared parent admission lock serializes this group's and the Gamma/Coins controllers' submissions. CER has its own root lock, so its admission is not serialized by that parent lock. Its four-worker limit and global cap 185 are handled by keeping our cap at 181, preserving a four-slot margin. Coordinate near the caps before changing these limits; do not describe all studies as sharing an atomic admission transaction. A live follow-up confirmed CER's frozen 612145f pin, max four, cap 185 and zero submissions alongside our zero-submission controllers. No CER source, settings or cap change is needed.

Commons retains frozen worker source 612145f, its three original training budgets, five seeds, algorithms, schedules and four seed-zero setup gates. Response workers use the newly published source, including the numerical and unused-replay-tail repairs, with an explicit 4096-episode selection option. No shared remote checkout or venv is updated. Workers request one CPU, 16 GB and at most seven days; controllers poll at 15-minute intervals.

For the final 85-worker combined report, merge the new 60-worker Commons root, the new 15-worker response root and the ten-profile corrected CX3 calibration root pltlf-equilibrium-calibration-v2-20261012. Do not merge the stopped original 75-worker root or the superseded raw CX3 pilot root. The merger rejects overlapping worker IDs. Failures require diagnosis before any resubmission.

Concrete worker pins, controller PIDs, dry-run checks and first admission states are recorded below. No scientific outcomes are claimed by this protocol note.

Active deployment¶

At 11:29 UTC, the handoff checked the shared submission lock, all 75 worker identities, training arguments and memory requests, the old controller's process identity, and its empty ledger. It stopped only controller 3443484, then launched both replacements in detached process sessions. The old campaign.json bytes were preserved, with SHA256 bec27eaea9fafe5f5e3cc971d9174566a9bf08ac57390c9dc21c281c865e0bc0. Its response-study-handoff.json records zero cancelled scheduler workers.

Responses use detached checkout /vol/gpudata/oja24-pltlf/worktrees/equilibrium-responses-cec65cf; Commons continues to use equilibrium-612145f. Both venv links resolve to the existing project environment. Dry runs recorded exactly fifteen and sixty workers, and neither changed the shared checkout or environment. The first live snapshots confirmed both controllers alive, no failed workers and queue backpressure above cap 181. These are operational checks, not scientific results.

All fifteen archived inputs have finite parameter arrays, and their local copies match the remote SHA256 hashes. IPPO records canonical monitor numbering; IQL-CER uses the recovered formula-keyed permutations. The durable preflight is results/strong-response-input-audit-2026-10-12.json. Both new roots retain deployment.json, campaign.json and submission-preview.txt; responses additionally retain input-manifest.json and the exact handoff.py. Metadata copies are available under the matching local optimisation roots.

The updated local watcher initially used PID 1768477. It monitors the separate Commons and response roots alongside Pursuit and corrected CX3 calibration. At first response admission it queues a startup/log inspection. Each completed response's protocol, runtime, result and paired success arrays are copied locally before the completion notice. Failure, stalled-ledger and repeated connection-error checks remain active; the deduplication ledger is retained. Read .cache/equilibrium-campaign-watch.pid for the current watcher PID.

Validation: the broad changed-path check passed 545 routine tests and exposed one stale Coins storage-path assertion introduced by the separate full-view export isolation. The assertion now checks canonical task identity independent of the storage prefix; all 96 focused threshold, submission and response verifier tests passed, followed by all twelve threshold tests through the required changed-path selector. Ruff and strict Zensical builds passed. The paper update was published as a2ee015, and its combined Coins clarification is 4957d9f. The 34-page LaTeX build passed and the changed main evaluation and appendix pages were visually inspected. Pending spatial outcomes are explicitly excluded from paper success counts.

Submission-slot bottleneck¶

The 12 October read-only capacity investigation found a separate EqEx rescreen feeder, PID 3402703, targeting 195 expanded user jobs against the partition's 200-job submission limit. It checks every 300 seconds and refills when at least five slots are available. Held array tasks count towards that limit and towards our admission ceilings.

The holds have an identified owner: EqEx's cpu_gate.py --adopt-held, PID 3525087, implements Omar's requested 30-CPU reserve by limiting counted EqEx/omsh/accord usage to 258 of 288 CPUs. This is documented in EqEx's docs/AI/experiments/experiment-launch-scripts.md and commits b70a02b and 65a9b4e. The feeder itself does not issue holds or releases. A subsequent probe found 190 expanded tasks: 54 running and 136 pending, including 122 held rs-seg20 tasks. The CER, Commons, independent response, Gamma and full-view Coins controllers were alive; CPU study workers were still unadmitted. These snapshots are operational evidence, not experiment failures.

Reserving CPUs does not reserve submission slots. The feeder's target can therefore keep all paper controllers above their existing 185/181/177 ceilings. Preserve those ceilings, count held tasks, and coordinate new feeding with the EqEx owner before changing admission policy. No EqEx job, hold, feeder, gate, source or paper-study limit was changed in this investigation. Omar confirmed the EqEx owner is a Claude session. The recommendation relayed to Omar is to reduce its total-user expanded-job target to about 165, counting held tasks, and let the existing queue drain naturally. This leaves submission room beneath the unchanged paper-study ceilings; it is a proposal for the owner, not an applied configuration change. Both Codex peers have been informed. Existing workers and holds stay with their owner.

A read-only follow-up at 12:18 UTC confirmed that the owner implemented the 165-task feeder target. Old PID 3402703 is gone; new PID 3538041 runs the same script from index 190. Its default is now TARGET=165, its initial environment has no TARGET override, and its log records the corrected restart index after already submitting indices through 189. Expanded tasks, including holds, remain counted. The queue still had 195 tasks, including 122 held tasks, so admission waits for natural drainage. CPU gate PID 3525087 and all paper-study ceilings remain unchanged. W1/W3 made no external job, hold or feeder change.

Omar subsequently requested agent communication without messages posted on his behalf in Codex chats. W1/W3's watcher now records events in the shared agent mailbox instead of queuing chat turns; its deduplication ledger is preserved. A shared file-based mailbox is available to independent peers. See agent-coordination-2026-10-12.md for locations, provenance and the absence of automatic session wake-up.

The local watcher remains active. Its latest successful CX3 probe recovered after two connection timeouts and again found the corrected calibration controller active, with the two preserved corrected pilots and two submitted remaining workers. Connection failures are kept separate from worker failures.

First admission and startup checks¶

Read-only checks at 15:14--15:19 UTC found 60 expanded tasks, no held tasks, and 262 allocated Slurm CPUs against the shared 288-CPU limit. Submission backpressure has eased; CPU capacity remains a separate constraint. The Gamma/Coins agent reports that EqEx's owner subsequently centralized its reserve in cluster_shares.json (30 CPUs and 35 submission slots), with a feeder target of 165. Preserve the existing study ceilings and count any future held tasks. Historical zero-held observations reported by that peer were withdrawn after correcting %R's parenthesized reason; its expanded totals and study admission counts were unaffected.

Both independent response workers (Coins IPPO seed 0, job 61773, and IQL-CER seed 0, job 61787) are running with the frozen cec65cf source. Their protocol records retain 4096 selection and 8192 held-out episodes. The first has saved its initial 50k warm-response checkpoint; validation and later searches are unfinished. Startup logs contain no exception. Four 200k Commons seed-zero controls are running from 612145f, also without startup exceptions. No response or Commons scientific result is complete.

Six scheduled Pursuit audits have completed: Classic IPPO seeds 0--4 and Capture IPPO seed 0. Their 256-episode consistency checks all agree with the version-2 exact values. Exact maximum gains reproduce the existing development findings: Classic 0.461099, 0.309801, 0.172691, 0.464001 and 0.602726; Capture 0.722596. None passes epsilon 0.05. These repeat the same archived profiles and do not add independent seeds to the paper denominators. Result/runtime copies are preserved in the local campaign's audit/ root. Two continuation workers are running and two further Pursuit workers are admitted/pending, within the original four-worker allowance.

Both CER setup pilots completed and four full Territory arms are running; the setup outputs are engineering checks, not scientific performance results. The fresh Gamma worker is running and full-view Coins has admitted its first worker. Corrected CX3 calibration retains two completed preserved pilots and two further active runs. All scientific sources, budgets, thresholds and existing concurrency limits remain unchanged.

The peers relayed Omar's desired 1--2-day completion target. Shared allocation planning is underway; no proposed concurrency increase has been applied. First-response checkpoint/validation timing is needed to assess the longest serial worker. At current small concurrency, the entire set of campaigns cannot be promised within 48 hours. Existing local million-step continuations took about 5.5--5.8 hours on two local logical CPUs; those are development runtime references, not a measured cluster ETA. Peer communication and acknowledgements use the shared mailbox; no user-chat turn is injected.

Capture 2 and paper scope¶

Retain Capture 2 and describe original joint-threshold infeasibility in the main evaluation remarks. The default two-player, horizon-100, slip-0.1, event-delayed instance has numerical guard minimax value 0.129895377488. An unrestricted initial-state 0.05-NE has guard payoff at most 0.179895377488, below its original threshold 0.3. The intruder threshold 0.5 is feasible. This does not justify extending the conclusion to Capture 3, different game parameters or equilibrium against a restricted learner class. The known-model reference stays outside model-free learner success counts. No threshold is changed: a joint-feasible replacement would be a substantive new benchmark.

Fixed existing-response witness pilot, 14 October¶

The five retained Territory/IPPO workers have fourteen or fifteen of their twenty-eight planned response checkpoints saved, with no held-out result yet. To resolve useful negative cases sooner, a separate evaluation freezes each seed's already-saved player-zero native-warm PPO response after 200,192 training actions. The response identity is declared before new held-out data; no further response or joint-policy training is performed. All five baseline hashes match the broad finite-checkpoint audit. The original searches, architectures, sources, schedules, objectives and budgets remain unchanged.

The pilot is seed zero, Slurm 62626, using one existing Pursuit verification slot under cap 181 and the shared admission lock. It was independently confirmed RUNNING and reached its baseline evaluation after allocated checks verified input hashes, finite parameters/optimizer state and native update counters. It evaluates baseline and fixed response on 8,192 paired fresh episode seeds, starting at 850503 plus one million times the profile seed; these are disjoint from the original validation and held-out blocks. The unchanged source is cec65cf. The other four fixed cases are declared but unsubmitted until the pilot's usefulness is assessed.

The declared Territory response envelope is 35 budget/profile cases: 10 ordinary-IPPO cases, the existing conservative 20-case CER envelope, and five fixed warm responses. New paired assessments use delta 0.025 over this combined envelope. Final aggregation must recompute the original paired-count intervals with the same 35-case envelope while preserving raw results and frozen worker protocols. These response comparisons are distinct from the 1,491 declared whole-profile candidates and add no whole-profile replacements.

A profitable deviation with a lower confidence endpoint above epsilon 0.05 can resolve instability and justify retiring that profile's remaining search after review. An unsuccessful fixed response leaves stability uncertified and the original search running. The read-only observer records progress and validated results without automatic submission or cancellation. The protocol and startup evidence preserves every input hash, source, seed block, allocation and deployment checksum. Private files are in .cache/territory-fixed-warm-witness-20261014/; no new performance conclusion is available at startup.

Combined response correction, 14 October¶

The two completed Territory/IPPO-CER searches have now been reassessed under the combined 35-case response envelope. All eight saved paired arrays, both frozen results and both protocols match the previously published hashes; baseline episode bits, gains and paired disagreement counts also agree. No training, evaluation or candidate selection was repeated. Thirty-two independent inversions of binomial tails verify the new confidence endpoints.

Seed one's 50k response gain remains 0.089355, with interval [0.070556, 0.107933]; its 200k maximum gain remains 0.119385, with interval [0.098464, 0.140013]. Both remain profitable-deviation witnesses at epsilon 0.05. Seed three's maximum gain remains 0.023682, with interval [0.015460, 0.031813] at either budget; this is still a finite-search result without an equilibrium certificate. The combined-family reassessment supersedes the older 20-case intervals for current aggregation while preserving the original report and worker files. Pending ordinary-IPPO and fixed-response results must use the same combined envelope.